Skip to content
be-human.ai
← Back to the journal

Context & memory

Company Glossary for AI Agents: Terms Agents Must Use

A controlled glossary of product, job and status terms keeps AI agents using the same words over time. Here is how to build, load and review one.

Be Human typographic cover with the title: Company Glossary for AI Agents: Terms Agents Must Use
Conceptual illustration created with AI · Be Human

A company glossary for AI agents is a short, owned list of the product, job and status terms your business uses, with one preferred label and one plain definition for each. When an agent reads that list before it writes, it keeps naming the same things the same way over weeks and months, instead of drifting into synonyms that break reports, routing and customer replies. This article explains which terms to include, how to write entries an agent can follow, and how to keep the glossary under review.

Key takeaways

  • Wording drift is a memory problem: an agent repeats the labels it sees most often in its recent context, so inconsistent notes produce inconsistent language.
  • Keep the glossary limited to terms that drive decisions, routing, status changes or customer-facing replies.
  • Every entry needs one preferred label, one plain definition, a list of banned synonyms and a named owner.
  • Load the glossary as reviewed memory that the agent reads first, and change it only through a logged approval step.
  • Test wording on a recurring schedule so drift shows up as a failed check, not as a confused customer.

Why agents drift in their wording

Drift rarely looks like an error. A single reply that calls a pending quote 'waiting for approval' is reasonable English, and nobody flags it. The trouble starts when that phrasing is copied into a summary, the summary becomes memory, and a later reply uses the new phrase again. After a few cycles the agent is using three labels for one status, and the tool that expects the exact value either rejects the update or files it under the wrong heading.

The causes are ordinary. Staff use different words in different channels. Older documents carry outdated product names. Summaries rephrase whatever they compress. A model has no built-in way to know that a site visit and an inspection are, in your company, two different jobs with different people and different paperwork. It only knows what is in front of it. A glossary gives it a fixed reference, so the meaning does not depend on whichever note happened to be retrieved last.

Which terms belong in the glossary

Start with three families. Product terms are the names of what you sell, as customers and staff actually say them, along with the names you want used in writing. Job terms name the units of work: intake, estimate, change order, shift, delivery, case, appointment. Status terms name the states a job or request can be in and the moves between them. A fourth, small list holds banned synonyms, the words that have caused trouble before.

Do not try to define the whole vocabulary of the company. A useful first glossary is often a few dozen entries, chosen because they appear in tools, invoices, schedules or messages to customers. Include a term when at least one of these is true: it changes a status, it decides who handles a request, it appears in a document or message a customer will read, or two people on your team already disagree about what it means. Leave out general words that carry no operational weight.

How to write an entry an agent can follow

An entry that works for a person often fails for an agent, because people fill gaps from experience and agents fill gaps with guesses. Each entry should state the preferred label exactly, give a definition of one or two sentences, say what the term is not, list the allowed values if it is a status, name the owner who decides changes, and show one correct usage and one incorrect usage.

Illustrative scenario: a small cleaning company uses 'recurring client' in its scheduling tool, 'regular' on its website and 'contract customer' in its invoices. An agent reading all three starts writing 'regular contract clients' in summaries. The fix is an entry that sets 'recurring client' as the only internal label, defines it as a customer with a scheduled service at a set interval, states that a one-time booking is not a recurring client, and bans 'regular' and 'contract customer' in new notes. The website copy can keep its own wording, but the agent's internal records do not.

Write definitions in the company's language, not in textbook language. If the field team calls a job a 'ticket' and the office calls it a 'case', record both and decide which one the agent uses in which context, rather than letting it choose on each occasion.

A glossary entry is a decision, written down so that it can be checked.

A step-by-step plan to build the first glossary

Build the first version in one focused working session with the people who own operations, then refine it using real output. The sequence below keeps the work small enough to finish.

  • Collect candidate words from the places where labels already live: status fields in your job or CRM tool, invoice descriptions, schedule titles and a sample of recent customer messages.
  • Group the candidates into product, job and status terms, and mark the ones that a system or a customer depends on.
  • For each concept, choose one preferred label and assign one owner who has the authority to change it.
  • Write the definition, the 'is not' line, the allowed values for statuses and the banned synonyms.
  • Check each status value against the exact values your software accepts, so the glossary and the tool cannot disagree.
  • Load the approved glossary into the agent's memory as a versioned, reviewed record, and cite the version in every change log entry.
  • Run a short set of test requests and read the outputs for banned words and invalid statuses before the agent writes anything to shared records.
  • Set a recurring review, owned by a named person, that adds terms which appeared in disputes and retires terms nobody uses.

Keep the first version short. Fifty precise entries are more useful than several hundred vague ones, because the agent reads the whole list on every task, and a crowded list invites mistakes in exactly the places that matter.

Enforcing the glossary when agents read and write

A glossary only works if it is present when the agent reads and when it writes. On the reading side, keep it short enough to sit at the top of every task context, and load only the entries relevant to the current job, following the same discipline you would use for any context budget, as described in the guide to context budgets. On the writing side, summaries and status updates should use preferred labels, and a simple check can flag banned synonyms before a note is saved as memory. Include these checks in the routine described in the memory audit guide so that glossary violations are caught alongside other memory problems.

Changes need a defined path. Who may edit an entry, and who may approve a new status value, should be set by role rather than by whoever happens to have access. The OWASP authorization guidance stresses checking permissions on every access instead of assuming them, which applies directly to memory that controls wording in customer-facing work. Every change should record what changed, who approved it and which agent versions were affected. When two memories disagree about a term, the approved glossary wins, and the conflict is logged for review, as explained in the guide to conflicting agent memories.

Be Human's managed approach treats this as part of operations. The glossary is designed with the business owner during the first engagement, loaded as reviewed memory with a named owner and a change log, and checked during recurring evaluations and permission reviews. Changes to terms pass through the same approval boundary as other memory changes, so the agent does not rewrite the company's vocabulary on its own.

Detecting drift before customers do

Drift is easiest to catch with a fixed set of test cases. Prepare a small collection of realistic requests: a customer asking whether a service is recurring, a staff member asking to move a job to its next status, an agent asked to summarize a phone call. For each one, write the expected labels and the forbidden words. Run the set after every glossary change, after model upgrades and on a recurring schedule. Treat a failed check as an incident with an owner, not as a curiosity.

The OpenAI evals guide describes the general pattern of building repeatable checks for model outputs, and that pattern fits wording well. Start with exact-match checks for status values, which are objective, and add a reviewed rubric for tone or clarity only where a person can judge it consistently. Keep the test set separate from the examples used to shape the agent, so that a passing score means the agent generalizes rather than memorizes.

Over time, the logs of wording failures show which terms are genuinely confusing. A term that fails repeatedly usually needs a clearer definition or a narrower scope, not just another ban. Feed those findings back into the glossary review, and the vocabulary becomes more precise rather than merely longer.

Take the free Be Human AI diagnostic to find the first task worth handing to an agent and see whether a glossary belongs in that first step.

Frequently asked questions

What is a company glossary for AI agents? A company glossary for AI agents is a short, reviewed list of the product, job and status terms your business uses, each with one preferred label, a plain definition and a list of banned synonyms. It is stored as memory that the agent reads before working, so the same concept is named the same way in notes, replies and system updates. A named person owns it, and changes pass through an approval step.

How many terms should a first glossary include? Start with terms that change a status, route work or appear in documents or messages that customers read, which usually produces a short list. A few dozen precise entries are more useful than hundreds of vague ones, because the agent reads the whole list on each task. Add terms when disputes or repeated wording failures show that a term needs a definition.

Does a glossary replace retrieval from company documents? No, the two do different jobs. The glossary fixes the meaning of words, while company documents hold the detailed policies and facts that the agent retrieves for each case. Where a term has a formal rule behind it, the glossary entry should point to the approved document that defines that rule.

How often should the glossary be reviewed? Review it on a fixed schedule owned by a named person, and also after any change to status values in your software, any new service line or any model upgrade that changes how the agent words things. Failed wording checks should trigger a review sooner. Each change should be logged with its version so that past outputs can be traced to the vocabulary that was in force.

Your next step

See where an AI agent fits in your company first.

Take the free Be Human AI diagnostic. A few questions about your work show which task to hand to an agent first, what it needs to know and where a person approves.

Be Human editorial guidance. Examples are illustrative unless explicitly attributed. Sources provide background; implementation choices depend on your workflow.

Less busywork. More room for you.

Bring us the work you wish you had time for.

Book a call to try a personalized AI agent free for 7 days ↗Or take the free AI diagnostic →